recipes for healthy food: AI Compute Doubling Every Six Months

Managing Editor's Note: Don't forget, our colleague Larry Benedict is hosting his Get Rich Slow event tomorrow at 8 p.m. ET…

That's where he's revealing one of his go-to strategies for snapping up quick profits on wild swings in the markets… over and over again… sometimes in a matter of just days or weeks.

And with all the volatility Larry is anticipating this year, there's a lot of opportunity with a strategy like that in your pocket.

If you haven't already done so, you can go here to add your name to Larry's guest list with one click.

AI Compute Doubling Every Six Months

By Jeff Brown, Editor,
The Bleeding Edge

Technology is advancing so quickly now…

And gaining a deep understanding of this acceleration can be difficult. It's a hard concept to fully grasp.

Just saying that it is happening, which is what most do, doesn't mean much.

There's no way to contextualize what "fast" or "very fast" or "super-fast" actually means.

And that's why I always try to provide context, data, and visualizations… so that we can grok the significance of this extraordinary period in time.

Exponential trends are particularly hard to quantify and contextualize.

Our human brains just aren't wired that way.

It takes a conscious effort to think exponentially.

This is one of the areas that we specialize in at Brownstone Research.

Wherever there is exponential growth, you can bet we're neck-deep in it – researching, providing our subscribers with the insights and investment recommendations to profit from these trends.

The underlying growth engine – currently fueling the exponential growth in technology and biotechnology today (which is now tech-driven) – is semiconductor technology and computational systems (servers, supercomputers, and hyperscale data centers).

If we don't deeply understand this, we can't possibly understand the trends.

The Truth About AI Compute

The chart below is a great visualization of what is happening right now.

It graphs the global artificial intelligence (AI) computing capacity since the first quarter of 2022.

And it shows how it has grown over time.

The chart has been normalized across five different companies' semiconductor solutions, with a unit of compute being equivalent to one NVIDIA H100e GPU.

Source: Epoch AI

I'm certain that every Bleeding Edge reader of mine will see the exponential curve in the chart above.

And it's also easy to see the inflection point around early 2024, when the curve starts to go vertical.

Classic example of exponential growth.

In this case, it is the exponential growth of global AI computing capacity.

And this capacity is directly related to breakthroughs in artificial intelligence.

Every increase in computational capacity results in exponential growth in technological breakthroughs.

That's why it is so hard to keep up with everything that is happening in tech/biotech.

And here's the key point that should put a pit in your stomach – that feeling of both excitement and anxiety.

What the above chart shows us is that global AI computing capacity is doubling every seven months.

And that doesn't truly capture the scope of what's occurring. Actually, it's now happening even faster than that.

The data in the chart above is only through the third quarter of 2025.

Public companies have yet to release their quarterly reports for Q4, which contains the information needed to extend the chart through the end of 2025.

And we're already halfway through January.

If I had to estimate, we're now doubling capacity every six months.

And that window in which computing capacity is doubled will get even narrower before all is said and done. But here's where the context is important…

Double the Capacity in Half the Time

Moore's Law, which became the reference example for exponential growth, measured the change in the number of transistors fitting into one unit of space.

Moore's Law accurately predicted a doubling every 18-24 months.

And along similar lines, we have seen related exponential growth in the computational power of semiconductors, and the exponential decline in the cost-per-unit of compute. (If you missed my discussion on this trend – and the rapid decline in intelligence-to-price ratio, I suggest you catch up right here in The Cost of Intelligence. It might just be the single most overlooked exponential trend happening right now.)

One year ago, global AI computing capacity was doubling at a pace of about every 10-12 months.

Think about that.

In just a year, the doubling in AI compute has gone from every 10-12 months… to roughly every six months.

And yes, by the end of the year, it will probably be happening every three or four months.

I can feel it in my gut right now. The implications…

And yet, that chart above doesn't even paint a complete picture.

It only highlights five companies' semiconductor platforms for AI:

NVIDIA (NVDA) with its GPUs, which make up about 60% of total compute
Google (GOOGL) with its TPU semiconductors
Amazon (AMZN) with its Tranium semiconductors
AMD (AMD) with its GPUs and inference semiconductors
Huawei (China-based), with its Ascend semiconductors

Notably absent from this list is everyone else. So, what about the rest?

Recommended Links

Tomorrow, Thursday, January 15, 8 p.m.
Get Rich Slow

"I've discovered how to collect ongoing double-digit gains from the increasingly wild swings in stocks and crypto. Here's how to tap into this new income stream." – Larry Benedict. Register for free here.
(When you click the link, your email address will automatically be added to Larry's guest list.)

Buy This Gold Stock Today

America is about to see a massive shift in how regular people buy and store gold. This year, the world's largest gold buyer is expected to launch a new way for everyday Americans to invest in gold with a simple tap on their phone – and it could go live in 2026. When it does, a tiny gold stock trading around $1.60 could explode. Gold would need to jump another $4,000 an ounce for you to double your money… But this stock only needs to rise $1.60 to do the same. Click here for the full story – plus four more gold stocks, Garrett Goggin, Founder of Golden Portfolio, believes could soar.

The Value of Inference

What about all of the other companies that are producing bleeding-edge semiconductors for AI applications?

Groq, for example, developed its Language Processing Unit (LPU) – optimized for high-performance inference – i.e., the running of AI applications.

Groq LPU | Source: Groq

Groq is not to be overlooked or "left out."

NVIDIA's GPU architectures are optimized for the training of large AI models, which makes them comparably inefficient for running AI applications.

This was the impetus for Groq's design in the first place.

Groq's LPUs have very low latency without sacrificing accuracy when running these AI applications.

And now we know how much Groq's technology is worth.

On Christmas Eve, NVIDIA announced that it was "acquiring" Groq for $20 billion.

The deal was structured as a $20 billion perpetual licensing agreement, combined with NVIDIA taking the most valuable Groq executives and employees as part of the "licensing deal." (I wrote about the incredible NVIDIA-Groq deal here.)

NVIDIA would have never gotten away with an outright acquisition – antitrust regulations would never allow for it – which is why it structured the deal the way that it did.

It simply left Groq as a shell of a company, offering cloud-based services using Groq's LPUs.

Or how about Cerebras and its Wafer Scale Engine (WSE-3)? It designed the largest semiconductor in the industry, utilizing an entire silicon wafer.

Cerebras Wafer Scale Engine 3 | Source: Cerebras

Cerebras took a different design approach than Groq, but its target was the same… low-latency inference applications. And unlike Groq, Cerebras has been gearing up for an IPO, which I believe will absolutely happen this year.

That is, unless another company steps up and acquires it first.

Ripe for Acquisition

SambaNova Systems is another major player in the AI inference industry, with its Reconfigurable Data Unit (RDU) designs.

While SambaNova also targets inference applications, it took a unique approach with its memory architecture, combining on-chip static random-access memory (SRAM) with high-bandwidth memory (HBM) and high-capacity double data rate (DDR) memory.

This gives SambaNova the ability to run multiple AI models in memory and switch from one to another in microseconds.

SambaNova SN40L RDU | Source: SambaNova

SambaNova attracted the interest of Intel, which has practically missed out on the entire boom in AI semiconductors.

Intel has been in advanced negotiations to acquire SambaNova for $1.6 billion, a price I believe is far too low for what SambaNova has built.

And we're just scratching the surface with these three companies alone.

There is also Tenstorrent, d-Matrix, and potentially the most interesting of them all – Neurophos – with its optical processing unit (OPU), bringing record-shattering energy efficiency and exaflop performance in a single server.

The key point here is that a lot more computational resources are being brought online outside of the big five players shown in the original chart.

And the pipeline of private semiconductors coming to market with optimized chips for one AI application or another is full of incredible potential.

Some will go public, many will be acquired, and all of them will accelerate the development and utilization of artificial intelligence.

Jeff